Integrity Review: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for CheXray Vision-Language Pre-training (ACM MM 2025)

Academic fraud report · Geng Detector

Summary

This report evaluates a 2025 ACM Multimedia paper (DOI: 10.1145/3746027.3755336) by Lihong Qiao et al. on chest X-ray vision-language pre-training. The overall verdict is highly suspicious, primarily driven by abnormal numerical patterns in Table 1. On the CheXpert (CXP) benchmark, the baseline MRM and the proposed model show identical AUC values to one decimal place across the 1% and 10% labeled-data splits (MRM: 88.5/88.5/88.7; Ours: 89.5/89.5/89.7), which is highly implausible given a tenfold change in training data. Additional concerns include suspiciously smooth incremental gains across other tables, the absence of standard deviations across runs, and reported COVIDx accuracy of 96.0% that appears excessive. Image-based checks (Western blot-style analysis, splicing detection, t-SNE hash comparison) could not be performed due to lack of pixel access. Findings should be treated as evidence-based suspicion requiring institutional verification.

Verdict

Highly suspicious (🟠). The strongest indicator is a numerical anomaly in Table 1 that is difficult to reconcile with normal deep-learning training variance. Image-level and code-level checks were not feasible from the available source.

Key findings

Suspiciously duplicated numbers in Table 1 (CXP linear probing). Identical AUCs at 1% and 10% labeled data for both baseline and proposed model:
MRM: 88.5 (1%), 88.5 (10%), 88.7 (100%)
Ours: 89.5 (1%), 89.5 (10%), 89.7 (100%)
Suspiciously smooth monotonic gains across Tables 1–6 with no reported standard deviations over multiple runs, which is atypical for rigorous medical imaging venues and may indicate selective reporting.
Striking COVIDx performance (96.0% ACC) that is claimed to surpass prior SOTA, contributing to a pattern of "too-perfect" results.
Image-reuse / splicing check not possible because only textual evidence was supplied; pixel-level analysis (background noise, splice-line scan, t-SNE hash) could not be executed.
Actionable next step: inspect the public repository at https://github.com/Felix1118/PADKB for hard-coded test outputs or fabricated evaluation scripts, and request raw training logs.

Evidence highlights

Location: Table 1, page 5 of the supplied PDF (The_Name_of_the_Title_Is_Hope__2_(2).pdf).
Quoted exact figures (CXP, ViT-based, linear classification):
MRM — 1%: 88.5, 10%: 88.5, 100%: 88.7
Ours — 1%: 89.5, 10%: 89.5, 100%: 89.7
DOI: 10.1145/3746027.3755336 (ACM MM '25).
Authors: Lihong Qiao, Shiyi Gao, Yucheng Shu, Bin Xiao, Weisheng Li, Xinbo Gao.
Implication: A tenfold increase in labeled data producing no change at one-decimal AUC precision, replicated across both a baseline and the proposed method, is a strong indicator of manually fabricated or copy-pasted values rather than measured results.

Notes

Confidence is high for the Table 1 numerical irregularity, moderate for the broader "selective reporting" concern, and not assessable for image-based manipulation.
This review is AI-assisted and intended for academic discussion only; final determination of misconduct requires investigation by the relevant institution, conference committee, or publisher.
No findings have been fabricated; any limitation in evidence (e.g., unavailable pixels for Figures 1–6) is explicitly stated rather than inferred.